Module silu_mul 

fused_silu_mul_split op-diff harness.

Input layout (matches the kernel API):

• gate_up: tokens × (2 * intermediate)
• For each token row: [gate ‖ up] concatenated Output:
• out: tokens × intermediate, where out[i,j] = silu(gate[i,j]) * up[i,j]

Structs

SiluMulOp